Before
Working late tonight, home a bit later
k
Danger 2/9Safe
True intentjust chatting
Needs: nothing
Open source · MIT · Android / Windows / macOS
Read them first,
then reply
Jev Chat Assistant is your chat copilot. It reads visible chats, judges intent and danger level, drafts replies, and fills the one you pick into the input box. Whether to send is up to you. Available for Android, Windows, and macOS.
It only fills the reply into the input box. It never sends on its own and never touches transfers, red packets or payments. Where does my chat text go?
Demo animation: the other person writes “You just don't get me.” The Jev overlay rates it danger 8/9, high risk; their true intent is checking whether you care, so it suggests acknowledging the feeling and holding off on substance. It then shows three ranked replies. Tapping Fill on the first one puts it in the input box; you press send yourself.
04
Tap Fill and the reply lands in the input box. The send button stays untouched: whether and how to send is your call.
Works in
- X (Twitter)
- Feishu / Lark (OCR)
- Any app (manual OCR)
- Web (planned)
What is Jev Chat Assistant?
Jev Chat Assistant is an open-source AI chat companion with Android, Windows and macOS versions. It helps you analyze messages, interpret intent and draft reply options. You review the wording and decide whether to send.
- Who it is for
- Individuals who want help choosing words, interpreting conversational tone or drafting a reply. Its assessments are suggestions, not certain conclusions.
- Platforms and feature scope
- Each platform is maintained separately. Accessibility capture, ML Kit OCR, permissions and chat-app compatibility on this page describe the Android version. See each desktop project for its own capabilities.
- Costs and requirements
- The Android project is free and open source under MIT. Model calls use your own API key and may incur provider charges. Desktop licenses and requirements are documented in their respective projects.
Project publisher: Finderchangchang. Find packages, platform documentation and updates on the GitHub project homepage.
01Read first
The same “k” can mean opposite things
Most assistants hand you a paragraph right away. Jev first works out what the message actually is, then decides how to answer.
Before
I forgot our anniversary. I'll make it up tomorrow
k
Danger 7/9Risky
True intentchecking whether you care
Needs an apology · Apologize first
Demo scenario; the chats are simulated.
Jev's decision model doesn't write. It picks from fixed options.
One request asks seven questions: true intent, danger level (0–9), what they need, whether to reply with substance now, what to do next and more. Every answer comes with probabilities and a confidence score, in about a second. A separate text model writes the replies.
- Choice
True intent
Checking you care · Venting · Asking for action · Wants an explanation · Casual chat · Topic closed
- Score
Danger level
0 is light chat, 9 is an open rupture; the overlay labels it safe, careful, risky or high risk
- Choice
What they need
An apology · Action · An explanation · Your care · Nothing
- Yes / no
Reply with substance now?
Should the next line carry real content; the overlay shows “give substance” or “hold off on substance”
- Choice
What to do next
Check the history · Apologize · Commit · Explain · Acknowledge · Say less · Make a plan
- Yes / no
Any subtext?
Meant literally, or a test, sarcasm or a hint
- Yes / no
Tension resolved?
They accepted and are joking again, or still waiting on you
Compared with other approaches
They all “help you reply.” The differences: how much they can read, whether they judge first, and who presses send.
| Capability | Jev Chat Assistant | Keyboard or in-app AI replies | General chatbots (copy and paste) | Auto-reply bots |
|---|---|---|---|---|
| Judges before writing | Yes | No | Partly | No |
| Reads any chat app | Yes | Partly | No | Partly |
| Non-invasive, no client changes | Yes | Yes | Yes | No |
| You press send | Yes | Yes | Yes | No |
| Notes + contact context | Yes | No | Partly | Partly |
| Pick your own API and model | Yes | No | No | Partly |
| Open source and free | Yes | No | No | Partly |
Based on the common, publicly available forms of each approach, not on any specific product.
02How it works
From a message on screen to a line in your input box
Only these steps in between. Solid steps run on your phone; dashed ones go to the endpoints you configure yourself.
On your phoneSent to your configured API
- 1
Read the screen
The accessibility service reads the open chat; if the text isn't there, it takes a screenshot and runs offline OCR
- 2
Jev judges
Intent · danger level 0–9 · reply with substance now or not
- 3
A text model drafts
Three casual replies: a safe follow-up, one with an action or commitment, and a short response. DeepSeek via OpenRouter by default
- 4
Jev ranks
Reorders them by fit and shows probabilities
- 5
Overlay shows
The read plus three replies, right next to the chat
- 6
Copy / fill
Never sends. You press send.
Notes and contact context
Keep notes and contact profiles (relationship, aliases, notes) on the phone. Pinned notes go with every request; the rest are picked up when their tags or titles match the current chat. Turn on history and the person's recent messages come along too. Replies are prompted to stay consistent with your notes and not to make up facts outside them. A contact can be linked across apps.
OCR when text is missing
When the accessibility tree has no message text, it takes a screenshot and runs ML Kit's offline Chinese OCR. No image is uploaded and no Google services are needed. In any app you can also trigger a one-off screen OCR from the floating-bubble menu.
Bring your own endpoints
Set the URL, key and model separately for the judge, reply and vision endpoints. Presets for OpenRouter, Bocha Jev, TypeSafe direct, DeepSeek official and Qwen-compatible APIs, each with a one-tap connection test. One OpenRouter key is enough to start.
One core, many apps
Judging, drafting, ranking and filling share one core. Supporting a new app takes one adapter of a few dozen lines; everything else is reused.
03Your data
It reads your chats, so here is exactly what happens
This is not a “collects nothing” app. Chat text does go to a model endpoint, but you are the one who sets that address. Everything else stays on your phone.
Your phone
- Reads the chat on screen
- Screenshots OCR'd on device, never uploaded
- Keys, notes and history (off by default) stay local
Only when it analyzes
The last 10 messages of this chat + your relationship note; matched notes and this person's history only if you turned them on
The model endpoint you configured
OpenRouter · Bocha Jev · TypeSafe · DeepSeek · Qwen · any OpenAI-compatible URL
The author's server
Not in the loop: the author runs no server and never handles your chats
- What it reads
- Only the conversation on screen, through Android's accessibility service. It doesn't modify any app's package, doesn't use their APIs or accounts, and doesn't read databases. You can limit it to an allowlist of chats, or turn off auto-analysis so it only reads when you tap.
- Where it goes
- Only at the moment of judging, and only to the endpoint you configured. The author runs no server and never handles your chats. What the provider does with the data is governed by its own privacy policy.
- Where it's stored
- Keys live only in the app's private storage. Chat text isn't written to disk or logs after sending. History is off by default and stays on the phone if you turn it on. OCR uses ML Kit's offline Chinese model; images are never uploaded.
- How to delete
- In Settings, “清空知识库与历史” (clear notes and history) removes them in one tap. Uninstalling wipes everything, including keys and settings. There is no cloud backup.
- What it never does
- It never sends messages on its own, never touches transfers, red packets or payments, and has no ads and no third-party analytics SDKs.
Full privacy policy (Chinese, with an English summary)
Every point can be checked against the source code on GitHub.
04Supported apps
Where it works today
Based on real-device tests; version numbers are the ones tested.
Androidv1.4
| App | Status | Notes |
|---|---|---|
| QQ Android | Full flow | Tested on 9.3.50 |
| X / Twitter DMs | Full flow | Tested on 12.25, Chinese interface |
| Feishu / Lark | OCR fallback | Verified on a real device; message text via offline OCR |
| Any other app | Manual | One-off screen OCR from the bubble menu |
| Web | Planned | No web version yet |
| iOS | Not yet | — |
05Get started
Get Jev for your platform
Choose the platform that fits you. Find installers, setup guides, and updates on the GitHub project page.
Androidv1.4
A chat copilot on your phone: reads messages, drafts replies, and fills the one you pick into the input box.
Android 11+ · arm64-v8a
Get Jev on GitHubWindowsv0.1.11
Captures your own chat window and runs OCR locally. No client changes, no database reads.
Windows 10 1903+ / 11
Get Jev on GitHubmacOSv0.6.0
Watches the screen and uses a small local model to judge intent and risk level, then drafts replies in a chosen style and ranks them locally.
Apple Silicon · macOS 13+
Get Jev on GitHubIf Jev helps you, consider starring the GitHub project to support its maintenance. 6.7k Star
Android setup
-
1
Install
Get the APK from the GitHub project page and install it. Android 11+, arm64-v8a devices only. Or use adb:
Get Jev on GitHubadb install -r jev-assistant-v1.4-release.apk
After upgrading from an older version, switch accessibility off and on again in system settings so the service rebinds.
-
2
Add a key
Settings → 接口 (Endpoints) → put an OpenRouter API key in the judge endpoint. Leave the others blank and they inherit it. Each endpoint has a one-tap connection test.
The judge endpoint also has a “Bocha Jev” preset, free for a limited time; get a key at jev.bocha.cn.
-
3
Grant permissions
Accessibility, display over other apps, autostart and unrestricted battery. On Xiaomi / HyperOS the last two are a must, or the bubble gets frozen.
- Accessibility service
- Display over other apps
- Autostart + unrestricted battery
Practical guides
Setup steps and examples you can refer to when using the project.
Set up Jev Chat Assistant: choose a platform and generate your first reply options
A practical Android setup guide covering model endpoints, accessibility, overlay permissions, costs and where chat data goes. Desktop editions have separate instructions.
How to review an AI-written reply before sending it
Use a clearly fictional conversation to check captured text, uncertain interpretations, invented facts and commitments before turning an AI candidate into your own message.
06FAQ
Before you install
Will it send messages for me?
No. It only fills the reply into the input box. It never sends anything on its own and never touches transfers, red packets or payments. Whether to send, and what to change first, is your call.
Is my chat uploaded?
Only at the moment of judging: the recent messages on screen and your relationship note go to the endpoint you configured (plus matched notes and contact notes if the knowledge base is on, and this person's recent history if history is on). Nothing is written to disk or logs afterwards, and the author runs no server and never handles it. Keys live only in the app's private storage. History is off by default and stays on the phone if you turn it on. OCR uses ML Kit's offline Chinese model; images are never uploaded. Full privacy policy (Chinese, with an English summary)
Which API keys do I need?
One OpenRouter API key is enough: put it in the judge endpoint and leave the reply and vision endpoints blank to inherit it. You can also set all three separately, with presets for OpenRouter, Bocha Jev, TypeSafe direct, DeepSeek official and Qwen-compatible APIs, each with a one-tap connection test. The Bocha Jev preset for the judge endpoint is free for a limited time; keys are at jev.bocha.cn.
Does it cost anything?
The app is free and open source. Model calls use your own API keys and are billed by the provider for what you use; the project never handles any money.
Do I need root?
No root and no modules. It is non-invasive: it only reads the conversation on screen, doesn't modify any app's package, and doesn't use their APIs or accounts.
iOS or desktop?
No iOS for now. The GitHub project page lists Android, Windows, and macOS; each platform is maintained independently. A web version is planned but doesn't exist yet.
The bubble disappeared, or it can't read messages?
Usually the Android skin froze the app in the background. Check that accessibility, display over other apps, autostart and unrestricted battery are all on, especially the last two on Xiaomi / HyperOS. Reinstalling resets the overlay permission, so turn it on again from the home-screen guide. Tapping anywhere in the chat usually brings it back.
Nothing happens after an upgrade?
Switch accessibility off and on again in system settings. Version 1.3 added screen capture, and the service has to rebind before it works.
Didn't find your answer? Open an issue or find community options on GitHub.
07Contact
Questions, ideas, or an app you want supported
Community and updates
Find the Chinese official account and group chat on the GitHub project page, where group QR codes are updated.
Community and updates on GitHubGitHub Issues
Bugs, an app it can't read, requests for a new app: Issues keep them trackable.
Open an issue